AI Apocalypse? Creating Complexity or Opportunity (Part 2 of 2)

Automating Cost Drivers Vs Deliberate Incentives for Success – by Michael Koriwchak, MD

The concern is a real one, and it reflects a broader tension in how AI is being deployed across industries: whether it is used primarily as a cost-containment weapon or as a value-creation tool. In insurance—particularly health insurance—that tension is especially acute because the “product” is not abstract efficiency, but access to care, financial protection, and, ultimately, patient outcomes.

What you’re describing could evolve into a kind of adversarial equilibrium: insurers deploying increasingly sophisticated AI systems to scrutinize, delay, or deny claims, while providers, patients, and intermediaries deploy their own AI tools to challenge those denials, optimize coding, and maximize reimbursement. The result is not a more efficient system, but a computational arms race—what you aptly call a “battle of the machines”—that increases administrative cost without improving care.

We have already seen early versions of this dynamic in the pre-AI era. Revenue cycle management systems, prior authorization workflows, and claims editing software have long been used by both payers and providers to optimize their respective financial positions. The administrative overhead in U.S. healthcare—often estimated at hundreds of billions of dollars annually—is, in part, the result of these competing optimization strategies. AI has the potential to dramatically accelerate and scale this dynamic.

If insurers primarily deploy AI to identify reasons not to pay—flagging claims for minor discrepancies, applying rigid interpretations of medical necessity, or automating denials at scale—then providers will respond in kind. They will use AI to preemptively optimize documentation, appeal denials automatically, and identify patterns in payer behavior to exploit weaknesses. Each side becomes more efficient at opposing the other, but the system as a whole becomes more complex, opaque, and expensive.

This is a classic case of local optimization leading to global inefficiency.

Avoiding this outcome requires a deliberate shift in how AI is positioned and governed within the insurance ecosystem. Instead of framing AI as a tool for denial or defense, it should be oriented toward reducing friction, improving accuracy, and aligning incentives across stakeholders.

On the payer side, this means using AI to improve the precision of payments rather than the frequency of denials. For example, AI could be used to:

• Ensure that claims are adjudicated correctly the first time, reducing the need for rework.
• Identify genuine fraud or abuse without casting an overly broad net that captures legitimate claims.
• Streamline prior authorization by approving routine, evidence-based care automatically and reserving human review for truly ambiguous cases.

On the provider and claim side, AI can similarly be used to:

• Improve documentation quality so that claims accurately reflect the care delivered.
• Reduce coding errors and variability, which are a major source of denials.
• Provide real-time feedback at the point of care, aligning clinical decisions with coverage policies before a claim is ever submitted.

The key is that both sides are using AI to reduce uncertainty and error, not to exploit it.

A critical concept here is “first-pass yield”—the percentage of claims that are paid correctly on initial submission. In a well-functioning system, this should be as high as possible. If AI is deployed collaboratively (or at least symmetrically with aligned incentives), it can significantly improve first-pass yield, reducing the need for appeals, resubmissions, and manual intervention. That is where the real economic value lies—not in denying more claims, but in avoiding the need to fight over them in the first place.

There is also an important role for transparency and standardization. One of the reasons adversarial dynamics emerge is that rules are अस्पष्ट, inconsistently applied, or subject to strategic interpretation. If AI systems are trained on opaque or proprietary criteria, they can amplify these inconsistencies. Conversely, if there is greater standardization of medical necessity criteria, coding practices, and adjudication rules—potentially supported by shared or auditable AI models—the scope for adversarial behavior diminishes.

Regulation and policy will likely play a decisive role. Just as earlier phases of the insurance industry required oversight to prevent abusive practices, AI-driven claims adjudication may require guardrails to ensure fairness, accountability, and explainability. For example, regulators could require that automated denials be explainable and subject to meaningful human review, or that AI systems be audited for bias and error rates.

Finally, there is a cultural dimension. Organizations tend to deploy technology in ways that reflect their underlying incentives and values. If insurers view AI primarily as a lever to reduce payouts, they will optimize for denial. If they view it as a tool to improve member experience and long-term outcomes, they will optimize for accuracy and trust. Similarly, providers can choose to use AI to “game” the system or to align more closely with evidence-based care and documentation.

The long-term sustainability of the system depends on avoiding a zero-sum mindset. A full-scale “battle of the machines” would not only be expensive; it would erode trust, increase patient frustration, and divert resources away from actual care. The more promising path is one in which AI reduces the need for conflict by making the system more predictable, transparent, and aligned.

In that sense, the lesson from earlier technological shifts applies here as well: the impact of AI will be determined less by what it can do, and more by how we choose to use it.